OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2026-01-15 12:15:51 +00:00
Code Issues Actions 52 Packages Projects Releases Wiki Activity

Revert "[profiler] enable CUPTI range profiler in build (#125685)"

Browse Source 
This reverts commit 2deea9e6e9.

Reverted https://github.com/pytorch/pytorch/pull/125685 on behalf of https://github.com/atalman due to Broke nightly ([comment](https://github.com/pytorch/pytorch/pull/125685#issuecomment-2103093237))
This commit is contained in:
PyTorch MergeBot
2024-05-09 17:28:01 +00:00
parent 26b942c4fc
commit 8fb3ff2a4e
2 changed files with 0 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
cmake/Dependencies.cmake
View File

@@ -1834,8 +1834,6 @@ if(USE_KINETO)
set(CUPTI_LIB_NAME "cupti.lib")
endif()
set(NVPERF_HOST_LIB_NAME "libnvperf_host.so")
find_library(CUPTI_LIBRARY_PATH ${CUPTI_LIB_NAME} PATHS
${CUDA_SOURCE_DIR}
${CUDA_SOURCE_DIR}/extras/CUPTI/lib64
@@ -1850,27 +1848,13 @@ if(USE_KINETO)
${CUDA_SOURCE_DIR}/include
NO_DEFAULT_PATH)
find_library(NVPERF_HOST_LIBRARY_PATH ${NVPERF_HOST_LIB_NAME} PATHS
${CUDA_SOURCE_DIR}
${CUDA_SOURCE_DIR}/lib
${CUDA_SOURCE_DIR}/lib64
${CUDA_SOURCE_DIR}/extras/CUPTI/lib64
NO_DEFAULT_PATH)
if(CUPTI_LIBRARY_PATH AND CUPTI_INCLUDE_DIR)
message(STATUS " CUPTI_INCLUDE_DIR = ${CUPTI_INCLUDE_DIR}")
set(CUDA_cupti_LIBRARY ${CUPTI_LIBRARY_PATH})
message(STATUS " CUDA_cupti_LIBRARY = ${CUDA_cupti_LIBRARY}")
# CUPTI Range Profiler requires the NVPerf library
# for configuring metrics
if(NVPERF_HOST_LIBRARY_PATH)
set(CUDA_nvperf_host_LIBRARY ${NVPERF_HOST_LIBRARY_PATH})
message(STATUS " CUDA_nvperf_host_LIBRARY = ${NVPERF_HOST_LIBRARY_PATH}")
endif()
message(STATUS "Found CUPTI")
set(LIBKINETO_NOCUPTI OFF CACHE STRING "" FORCE)
# I've only tested this sanity check on Linux; if someone
# runs into this bug on another platform feel free to
# generalize it accordingly

42
test/profiler/test_profiler.py
View File

@@ -699,48 +699,6 @@ class TestProfiler(TestCase):
if torch.cuda.is_available():
check_metrics(stats, "device_memory_usage", deallocs=["[memory]"])
@unittest.skipIf(not kineto_available(), "Kineto is required")
@unittest.skipIf(not torch.cuda.is_available(), "CUDA is required")
def test_kineto_cupti_range_profiler(self):
"""CUPTI provides a newer Profiling API from CUDA 10.0 that enables measuring
performance events for the GPU. This is supported as an experimental pytorch profiler feature.
Read more here https://docs.nvidia.com/cupti/r_main.html#r_profiler.
"""
exp_config = _ExperimentalConfig(
profiler_metrics=[
# Metrics list at https://docs.nvidia.com/cupti/r_main.html#r_profiler
# or use kineto__tensor_core_insts, kineto__cuda_core_flops
"kineto__tensor_core_insts",
"dram__bytes_read.sum",
"dram__bytes_write.sum",
],
profiler_measure_per_kernel=True,
)
with _profile(
use_cuda=True, use_kineto=True, experimental_config=exp_config
) as p:
self.payload(use_cuda=True)
def check_trace(fname):
with open(fname) as f:
trace = json.load(f)
self.assertTrue("traceEvents" in trace)
events = trace["traceEvents"]
found_cupti_profiler_events = False
for evt in events:
self.assertTrue("name" in evt)
if "__cupti_profiler__" in evt["name"]:
found_cupti_profiler_events = True
# PyTorch OSS CI runs in docker containers where the Range Profiler
# does not have sufficient privilege level (CUPTI_ERROR_INSUFFICIENT_PRIVILEGES).
# We can check that the profiler does not crash the job and the trace is not
# malformed, however do not check the actual presence of data.
self.assertTrue(1 or found_cupti_profiler_events)
with TemporaryFileName(mode="w+") as fname:
p.export_chrome_trace(fname)
check_trace(fname)
@unittest.skipIf(
IS_JETSON, "Jetson has a guard against OOM since host and gpu memory are shared"
)